AI teams can now be trained to stay in their assigned roles — cutting the defection rate from 46% to 8%

Researchers built a measurement system that quantifies how well individual AI agents stick to their assigned job in a multi-agent team, then used that measurement to fine-tune the agents during training. In practice, this means you can now train an AI team where each member actually does what you told it to do instead of wandering into another member's job.

Until now, when you set up multiple AI agents to collaborate on a task, they would routinely ignore their role boundaries and start doing each other's work — a failure mode that tanked overall performance. This paper shows you can measure that drift precisely and eliminate most of it through lightweight retraining. That matters because it's the difference between multi-agent systems that are brittle and unpredictable and ones that actually scale to real work.